When AI Makes One Stage Abundant, the Moat Moves to the Bottleneck

Discovered drug candidates roughly doubled over the past decade. Approvals stayed flat at about 50 a year. In 2024, FDA’s CDER cleared 50 novel drugs, the same neighborhood it has occupied for years. That divergence is the entire argument of Ben Liu’s essay “Uneven Frontiers,” published by a16z in June 2026, and it is the cleanest case study yet for a pattern every operator should be tracking.

Liu is co-founder and CEO of Formation Bio, an a16z-backed AI-native pharma company, so the conclusion he reaches is also the conclusion his company is built to monetize. Read his data with that incentive in view. The mechanics he describes, though, hold up independent of who profits from them.

Two Frontiers Moving at Different Speeds

Liu’s framing splits drug development into stages and asks which ones AI accelerates. Discovery and molecular design behave like software. They run on data, compute, and iteration, and they improve on the curve that everything digital improves on. Clinical development behaves like the physical world. You recruit human patients, you dose them, you wait years for endpoints to read out.

The numbers show what happens when one frontier races ahead of the other. Drugs generated per attractive biological target rose from two or three a decade ago to seven or eight today, sometimes past fifteen, with Liu projecting toward twenty, thirty, even fifty against the most attractive targets. In 2025, more than 100 programs each chased PD-1 and the GLP-1s. DelveInsight counted 200-plus programs against PD-1/PD-L1; IQVIA counted 193 obesity and GLP-1 assets in October 2025.

So the funnel inlet is flooding. The funnel outlet is the same diameter it has been for years. Liu’s sentence for this: “In a world of drug discovery abundance, the scarce capability is no longer making drugs, but rather, knowing which drugs are worth developing.”

Why Clinical Development Does Not Compress

The slow stage stays slow for reasons that compute cannot touch. Phase 2 trials take two to four years to read out. Distinguishing a true efficacy edge from noise can take a decade or more. Baseline Phase 2 probability of success sits around 30 percent, which means most of what enters the expensive stage fails there.

Liu lays out a sequence worth memorizing: “Drug discovery and design will improve first, toxicity prediction will improve next, and clinical efficacy will be last.” Each stage hands more uncertainty to the one after it, and the last stage is the one bounded by biology and regulation rather than by model quality. His blunt version: “As long as regulators require prospective human clinical trials, clinical development will remain the bottleneck. The limiting factor is not just computation; it is the irreducible time required to test medicines in humans.”

That word, irreducible, is the load-bearing claim. A faster model does not shorten a three-year endpoint. More candidates entering trials does not make any single trial conclude sooner. Abundance upstream raises the cost of being wrong downstream, because now you can afford to run far more shots than the bottleneck can absorb, and choosing badly wastes the scarce resource.

The Moat Migrates to the Constraint

Here is the operating consequence Liu draws. When the discovery stage commoditizes, the durable value stops living inside the discovery models. It migrates to the stage that did not speed up. The company that wins is not the one with the best generator of molecules. It is the one that decides, better than anyone else, which molecules deserve a slot in the bottleneck.

That is selection capability, trial design, biomarker strategy, and the judgment to kill a program before it consumes four years and a fortune. None of it is the part AI made cheap. All of it is the part AI made more valuable by flooding the input side.

We have made the abstract version of this argument before. Capability is a commodity; orchestration is the moat traced how raw model power gets cheap while the system that directs it stays scarce. The domain-expertise tax showed that in spatial biology the bottleneck was never the model, it was the expert judgment required to use it. And the adoption deficit named the same shape at macro scale: substrate capacity outran the organizational capacity to absorb it. Biopharma is the physical, regulated, data-rich proof of the pattern. The constraint here is not organizational habit. It is human biology and prospective trial law, which is about as hard a bottleneck as exists.

The General Rule

Strip out the pharma vocabulary and the rule is portable. AI rarely transforms a value chain evenly. It commoditizes the stages that look like software first, leaving the physical, regulated, or trust-bound stages roughly where they were. Value does not stay where it used to sit. It pools at whatever AI cannot accelerate.

Legal discovery gets cheap; the courtroom standard of proof does not move. Marketing copy gets abundant; the brand-trust that takes years to build does not. Code generation gets cheap; the production incident at 3 a.m. still needs an accountable human. In each case the abundant stage stops being the place that captures value, and the slow stage becomes the asset, precisely because everyone now has the fast stage and almost no one has solved the slow one.

This is Liu’s argument, and it is also a self-interested one. Formation Bio is positioned to sell the bottleneck-side capability he says is scarce. The incentive does not make him wrong. It does mean you should test the claim against your own value chain rather than import his conclusion wholesale.

Do This Now

Map your value chain into stages and mark each one. Which stage is AI making abundant this year, and which stage stays bound by something physical, regulated, or trust-dependent that no model shortens. The abundant stage is about to stop being your differentiator, because your competitors are getting it too. The bound stage is where your durable advantage will sit, and it is probably underfunded today because it is the slow, unglamorous, expensive part nobody wants to own.

Move resources toward the bottleneck before the abundance arrives, not after. By the time the fast stage is obviously commoditized, the firms that built selection and judgment around the slow stage will already own the returns. In pharma that means knowing which drug to develop. In your business it means knowing which of the many cheap outputs is the one worth the scarce, slow, expensive next step. Find that stage. Build there.

---

This analysis synthesizes Uneven Frontiers (Ben Liu, Formation Bio, published by a16z, June 2026).

Victorino Group helps leaders find where AI commoditizes their value chain and where the durable moat actually sits. Let’s talk.